“With a new decentralized training algorithm, we fine-tuned GPT-J (6B) on 3.53 billion tokens, resulting in…